Migrant Self-Selection: Anthropometric Evidence from the MassMigration of Italians to the United States, 1907–1925∗

Yannay Spitzeryannay.spitzer@brown.edu

Brown University

Ariell Zimranariell.zimran@u.northwestern.edu

Northwestern University

August 8, 2014

Abstract

Are migrants positively or negatively self-selected from within their populations of origin? We studythis fundamental and persistent question of the economics of migration using data on one of the largestflows of free migration ever—that of Italians to the United States between 1907 and 1925. We exploitnever-before-used stature data in the Ellis Island arrival records—from which we transcribed the heightsand other personal information of a random sample of 50,000 Italian passengers—combined with Italianprovince-birth cohort height distributions and our own geo-matching of millions of Italian passengers totheir places of origin in order to construct a novel data set for our analysis. Relying on the well-establishedrelationship between population average stature and living standards, we quantify migrant self-selectionby comparing the heights of migrants to the height distributions of their respective birth cohorts intheir provinces of origin. Our analysis reveals opposite patterns of self-selection across and within Italianprovinces. Italian migrants were shorter, on average, than all Italians of the same birth cohort, suggestingnegative self-selection on the national level. However, when compared only to the distribution of staturein their own provinces of origin, we find that Italian passengers were, on average, taller, indicating positiveself-selection on the local level. Moreover, we find that the degree of self-selection from a province andbirth cohort was decreasing in its average stature, suggesting that less-developed province-cohorts, whereliquidity constraints to migration were more likely to bind, provided relatively higher quality migrants.The findings of this research demonstrate the importance of distinguishing between self-selection froma country as a whole and self-selection from within a particular sub-national region. Comparisons ofmigrants to their national-level origins, which are the norm in the literature on migrant self-selection,may fail to capture a significant portion of the self-selection occurring within a group of potential migrantsfrom a particular sub-national region.

∗The most recent version of this paper can be found at http://aez.econ.northwestern.edu/spitzer_zimran_italian_stature.pdf. The results in this paper are preliminary and may be affected by ongoing research and data transcription. Pleasecontact the authors before citing or circulating this paper. A previous version of this paper circulated under the title “Self-Selection of Immigrants on the Basis of Living Standards: Evidence from the Stature of Italian Immigrants at Ellis Island,1907–1925.”

1

mailto:yannay.spitzer@brown.edumailto:ariell.zimran@u.northwestern.eduhttp://aez.econ.northwestern.edu/spitzer_zimran_italian_stature.pdfhttp://aez.econ.northwestern.edu/spitzer_zimran_italian_stature.pdf

Acknowledgements We are indebted to Joel Mokyr, Joseph Ferrie, Igal Hendel, and Aviv Nevo forencouragement and guidance, and to the Northwestern University Economics Department’s Eisner Fund,the Northwestern University Center for Economic History and an Exploratory Travel and Data Award fromthe Economic History Association for financial support. We are grateful to Peg Zitko and the Statue ofLiberty-Ellis Island Foundation for providing the Ellis Island arrival records data, to Brian A’Hearn, FrancoPeracchi, and Giovanni Vecchi for sharing their computed moments of Italian stature distributions, andto Jordi Martí-Henneberg for sharing historical GIS files of Italy. We also thank Luigi Guiso, TimothyHatton, Seema Jayachandran, John Komlos, Lee Lockwood, Andrea Matranga, Paola Sapienza, MarianSmith, Richard Steckel, and Zachary Ward for helpful suggestions and insightful comments. Thanks are alsodue to Roy Mill for giving us access to the dEntry transcription system and for investing considerable time andenergy adjusting it to our needs; to Daniel Bird, Maureen Craig, Aanchal Jain, and Anand Krishnamurthyfor helpful discussions; to seminar participants at Northwestern University and conference participants atthe 2014 Warwick Economics PhD Conference and the Cliometrics Conference; and to Joshua Picache,Kris Angelo Belino, Abmelaine Pastores, Chermilyn Sarmiento, and Mary Rose Manlapaz for excellenttranscription. All errors are our own.

2

1 Introduction

[A]lthough drawn from classes low in the economic scale, the new immigrants as a rule are thestrongest, the most enterprising, and the best of their class . . . .

(The Dillingham Commission, US Congress, 1911, p. 24)

Between 1892 and 1925, nearly four million Italians immigrated to the United States—the largest sin-

gle flow during the Age of Mass Migration (Ferenczi and Wilcox, 1929, Tables 2–3, pp. 384–393). This

phenomenon, part of a general contemporaneous trend of growth in migration to the United States from

other southern and eastern European countries, sparked a debate over the policy of nearly total openness

of the United States to immigration (Goldin, 1994). Public debate focused primarily on the “quality” of

the southern and eastern European migrants.1 Groups favoring the restriction of immigration warned of a

decline in the quality of immigrants, arguing that these immigrants, unlike those arriving en masse from

northern and western Europe in prior decades, represented the poor, incapable, uneducated, unskilled, and

criminal elements of their origin countries; that is, that they were negatively selected from within their

environments of origin. In the late 1910s and early 1920s, after decades of agitation, such sentiment finally

prevailed with the passage of sweeping immigration restrictions, culminating in the Immigration Act of 1924,

which effectively ended unfettered large-scale immigration from Italy and other countries in the European

periphery (Hatton and Williamson, 1998, ch. 9).

Despite these allegations and the abundance of research, both modern and contemporary, that they

precipitated (e.g., Gomellini and Ó Gráda, 2013; Hall, 1904; Stolz and Baten, 2012; US Congress, 1911), the

question of whether migrants during the Age of Mass Migration were positively or negatively self-selected

remains unresolved.2 Even in the modern context, determining the nature and causes of migrant self-selection

remains at the forefront of research in the economics of migration (Borjas, 1987; Chiquiar and Hanson, 2005;

Fernández-Huertas Moraga, 2013; McKenzie and Rapoport, 2010), and is crucial in understanding the effects

of migration on the source and host economies (c.f., Biavaschi and Elsner, 2013). If, for example, migrants

are positively self-selected from within their populations of origin, then emigration, by disproportionately

leading to the exit of more productive individuals from the sending economy, may harm it (Bhagwati, 1976;

Di Maria and Stryszowski, 2009; Docquier and Rapoport, 2012; Mattoo, Neagu, and Özden, 2008; Todaro,

1We use the term “quality” here to refer to any traits that affect an individual’s productivity. Examples include education,skill, health, wealth, and intelligence. Proponents of immigration restriction in the early 20th century had an even broaderdefinition, arguing, for example, that these new immigrants were more likely to be involved in criminal activity, or lacked ahistory of self-governance that would be crucial to their assimilation in the United States.

2Abramitzky, Boustan, and Eriksson (2012, 2013) also study the self-selection of migrants in the Age of Mass Migration,focusing on Norwegians.

3

1996). Conversely, the receiving economy may benefit from the influx of these productive individuals. If

migrants are negatively self-selected, the opposite may occur.

However, empirical answers regarding migrant self-selection remain elusive, primarily due to a number

of data limitations that make direct comparisons between migrants and the population at risk for migra-

tion difficult, if not impossible in some contexts. In particular, a lack of representative data on the source

population often confounds efforts to quantify migrant self-selection. Even when comparison data are avail-

able, they generally cannot be disaggregated to geographic levels below the country of origin, raising the

possibility that the nature of self-selection from within source populations is obscured by composition effects

across sub-national units. Moreover, most data on migrant quality are observed only after arrival in the host

country, raising the possibility that they do not reflect the pre-migration characteristics of migrants. In most

of the few cases in which these issues can be overcome, the measure typically used to compare migrants to

stayers is occupation, which, although informative regarding individual skill and human capital, comprises

only a rough measure of a migrant’s economic capability, reflecting only limited aspects of it.

In the present research we study self-selection into migration using stature to measure migrant quality.

This approach is grounded in a large body of research, which has established that the average stature of a

large group is indicative of the group’s average economic capability—an amalgamation of many of its facets,

such as skill (Komlos, 1990), education (Case, Paxson, and Islam, 2009), income, (Deaton, 2007; Persico,

Postlewaite, and Silverman, 2004), wealth (Floud, Wachter, and Gregory, 1990), health (Fogel, 1986; Steckel,

1995), childhood environment (Bailey, Hatton, and Inwood, 2014), and cognitive ability (Case and Paxson,

2008), that all determine a prospective migrant’s contribution to his home economy, and his labor market

outcomes in the host economy. Stature is thus, at an aggregate level, a proxy for economic capability and

productivity in a broader sense than are other commonly used measures, such as occupation-implied skill.

Moreover, adult stature is fixed for those in a relatively broad age range and, for such individuals, is unaffected

by migration. The premise of this paper is that migrant self-selection can be quantified by comparing the

average stature of migrants to that of the populations of origin. If, for example, the population of migrants is

taller, on average, than the overall population in the sending economy, then it can be deduced that migrants

are more economically productive than non-migrants on average.

Applying this approach to the Italian migration to the United States enables a very reliable comparison

of the migrant population to the source population, as the stature distribution for the source population

of Italian adult males is known.3 Moreover, this distribution is known at a geographically disaggregated

3In most countries, military height data are only available for a self-selected group of individuals choosing to join themilitary. In Italy, however, all males were required to be measured by the military. The resulting data were collected by

4

level, enabling us to avoid obstacles stemming from the fact that the Italian migration was composed of

individuals originating in many heterogeneous provinces, and to explore the relationship between provincial

characteristics and different features of migration from each province. Studying a historical episode of

migration also carries many advantages that are unavailable in the study of modern migration. This approach

effectively avoids difficulties created by the fact that modern migratory flows are censored by restrictive

immigration policies, and are thus not representative of the latent supply of those willing and able to migrate.

Studying historical migration in which such barriers did not exist, allows scholars to cleanly identify migrant

self-selection at the source, to learn the mechanisms that determine the nature of migrant self-selection, and

to use these insights to make inferences regarding the effects of changes in migration policy on the quantity

and the quality of migrants.

In order to perform this analysis, we constructed a novel data set consisting of the stature, place of

origin, and other personal information of Italian passengers from the Ellis Island arrival records database.

First, we created a geolocation algorithm to assign each of the nearly five million passengers in the Ellis

Island database to his province of origin based on his reported last place of residence. Next, we randomly

sampled approximately 50,000 Italian passengers arriving between 1907 (when information on stature was

first collected on manifests of immigrants arriving in the United States) and 1925 (when the restrictions

of the Immigration Act of 1924 entered into force), and transcribed their stature and other information

regarding the nature of their voyage to the United States. We then compared the heights of migrants to the

distributions of Italian stature gleaned by A’Hearn, Peracchi, and Vecchi (2009) and A’Hearn and Vecchi

(2011) from Italian military records covering nearly all Italian males at conscription age.

This analysis reveals opposite patterns of self-selection across and within Italian provinces. Italians

passing through Ellis Island were shorter, on average, than all Italians of the same birth cohort, providing

evidence of negative self-selection on the national level. However, when compared only to the distribution

of stature in their own provinces of origin, we find evidence that Italian passengers were, on average, taller,

indicating positive self-selection on the local level. The difference between these two findings is driven by

positive self-selection within southern provinces, which were the origins of a disproportionately large share of

migrants, and in which the average stature was below the national average for Italy. Moreover, we find that

immigrants from northern Italy tended to be negatively self-selected from within their provinces of origin—

the opposite of their southern compatriots. Moreover, the degree of positive within-province self-selection of

immigrants arriving in the United States after 1917 was far greater than that of immigrants arriving in the

A’Hearn, Peracchi, and Vecchi (2009) and A’Hearn and Vecchi (2011), and made available to us.

5

pre-1917 period.

We further investigate what factors determine the degree of self-selection from within provinces, thus

providing a test of three major theories of migrant self-selection—relative inequality, liquidity constraints,

and network connections.4 We find that the degree of migrant self-selection was decreasing in the level of

development of the province of origin (as measured by its average stature), indicating that immigrants from

relatively less-developed environments were, on average, of higher quality relative to their provinces and

birth cohorts of origin than those from relatively more developed environments. We also find evidence that

migrants who were able to finance their own passage were more positively self-selected, on average. These

results are consistent with theories that predict positive self-selection due to the need to overcome liquidity

constraints to migration. We also find that individuals who migrated to join an immediate family member

were, on average, shorter than those who did not. This finding is consistent with the notion that chain

migration particularly helpful for lower quality migrants in overcoming liquidity constraints to migration.

We do not find any robust and statistically significant evidence supporting theories that hold that the nature

of migrant self-selection is determined by the relative inequality of the sending and receiving countries.

Although taken from the report of the anti-immigration Dillingham Commission, the epigraph to this

paper, like our results, demonstrates the importance of distinguishing between self-selection from a country

as a whole and self-selection from within a particular sub-national region, and of conceding that the two levels

of self-selection may be qualitatively different. Comparisons of migrants to their national-level origins, which

are the norm in the literature on migrant self-selection (e.g., Chiquiar and Hanson, 2005; Stolz and Baten,

2012), may fail to capture a significant portion of the self-selection occurring within a group of potential

migrants from a particular sub-national region.

The remainder of the paper proceeds as follows. Section 2 provides the relevant historical and economic

background for this study. Section 3 discusses the data construction process, and provides summary statistics

for the data set used in this study. Section 4 presents the main results, which are interpreted in section

5. Section 6 evaluates various theories of migrant self-selection. Section 7 discusses possible threats to

identification. Section 8 concludes.

4These theories are discussed in more detail in section 2.2.

6

2 Background

The issue of migrant self-selection, particularly from Italy in the early 20th century, but also in the context

of modern migration, has been studied extensively by economists, modern and historical. In this section, we

provide background on the mass migration of Italians in the early 20th century, as well as on the body of

economic knowledge on the issue of migrant self-selection.

2.1 Historical Background

At the beginning of the 20th century, Italy lagged behind most other western European countries in terms

of nearly every economic indicator. As shown in Figure 1, real wages were low, less than half their level

in Britain (O’Rourke, 1997). Moreover, Italy’s industrial production lagged that of its neighbors (Ciccarelli

and Fenoaltea, 2013), and malaria and other diseases were endemic, particularly in the south (Foerster,

1919). As a result, living standards in Italy, measured by average stature, fell short of those of most other

European countries, as depicted in Figure 2. These poor economic conditions spurred many Italians to leave

their home to seek opportunity elsewhere. Such was the strength of this incentive that by the turn of the

century, Italy had become the largest source of migrants to the United States, displacing such countries as

Ireland, Great Britain, Sweden, and Norway. Moreover, as depicted in Figure 3, Italy led Europe in terms

of relative migration, with the highest rates of emigration per capita of any European country in the period

1900–1913.5

In some ways, the Italian migration was typical of the Age of Mass Migration: migrants were mostly

young, unskilled, and male; but in other ways, Italians were distinct from other migrants. First, they tended

to distribute themselves between several destination countries, primarily the United States, Argentina, and

Brazil. Between 1886 and 1895, nearly 75 percent of Italians traveling to the Americas went to Argentina or

Brazil, with the remainder going to the United States. By the period 1906–1915, the United States became

the lead destination for Italians, drawing more than twice the numbers of Brazil and Argentina combined.

There was also considerable (mostly seasonal) migration to other European countries of a magnitude rivaling

that of the flow to the United States (Hatton and Williamson, 1998, Table 6.1, p. 101). Second, roughly three-

quarters of migrants were male (Hatton and Williamson, 1998, p. 102), a gender imbalance exceeding that

of almost any other group. Finally, Italians were, more than any other group of migrants, likely to migrate

temporarily rather than to remain abroad permanently (Gomellini and Ó Gráda, 2013). The canonical

5If we include sub-national ethnic groups, however, Russian Jews were more likely to emigrate than Italians (see Spitzer,2013, 2014, for more details).

7

example is the tendency of these so-called “birds of passage” to exploit the seasonal differences between Italy

and South America, traveling between the two in order to participate in both harvests (Foerster, 1919; Hatton

and Williamson, 1998). Return migration also occurred from the United States; as the annual nominal wage

in the United States was nearly five times that in Italy, and the cost of round trip passage would consume

only 20 percent of those earnings, many Italians would travel to the United States to work, and then return or

remit the money to their families in Italy (Gomellini and Ó Gráda, 2013). It has been estimated (Bandiera,

Rasul, and Viarengo, 2013, Table 4, p. 37) that as many as 80 percent of Italian migrants to the United

States eventually returned to Italy.

Discussing emigration from Italy as a whole obscures the considerable variation in emigration rates and

the general patterns of migration across Italian provinces (Hatton and Williamson, 1998, p. 106). Whereas

most Italian emigrants in the 1880s were from the north (Gomellini and Ó Gráda, 2013), the poorer south

had taken the lead in terms of emigration rates by 1900 (Ferenczi and Wilcox, 1929, Table 10, pp. 432–443;

Hatton and Williamson, 1998, Table 6.4, p. 107). The greater emigration rates from the south were driven

primarily by the fact that the south was much poorer than the north. As a primarily rural agricultural

economy characterized by lower real wages (Hatton and Williamson, 1998, pp. 115–116), southerners had

a stronger push to emigrate than did northerners. In addition, the north’s relative proximity to major

European labor markets caused many northerners to migrate within Europe rather than overseas. Thus,

the mix of Italian immigrants to the United States was primarily southern, and therefore drawn from the

relatively poorer portions of the country.

Americans were aware that the bulk of Italian immigration to the United States increasingly originated

in the poorer south, and many were displeased with the new growth of Italian immigration in the early

20th century. Many of those opposed to a continued openness to immigration felt that malicious forces

were at work in Europe to transfer the least desirable elements of the population of Europe to the United

States (Commissioner-General of Immigration, 1903; Hall, 1904), citing the immigrants’ lack of skill, and

(as perceived by anti-immigration activists) lack of mental and physical fortitude as evidence. Writing at

the height of the migration, the Commissioner-General of Immigration (1903, p. 70) asserted that

The great bulk of the present immigration proceeds from Italy, Austria, and Russia, and, further-more, from some of the most undesirable sources of population of those countries. No one wouldobject to the better classes of Italians, Austrians, and Russians coming here in large numbers;but the point is that such better element does not come.

Notably, claims of such negative selection were not seriously disputed (c.f., Douglas, 1919). Even ad-

8

vocates of continued openness to immigration accepted them, arguing that better measures ought to have

been taken to prevent the dependent and criminal elements from entering the United States (Brandenburg,

1904), that the United States’s tradition as a haven for immigrants was worth maintaining, and that the

immigrants would eventually converge (even physically) to American standards through their time in the

United States (Boas, 1911).

Nativist concerns led to the formation of the Dillingham Commission (US Congress, 1911), which was

charged with investigating the nature and effects of the mass immigration. After collection and analysis of

considerable data, the commission enumerated in great detail the negative characteristics of the immigrants,

ranging from their poor living conditions to their lack of education and skill, eventually concluding that

immigration restriction was necessary in order to protect the “national character.” These restrictions culmi-

nated in the literacy test of 1917 and finally in the quotas of the Immigration Act of 1924, which brought

an end to mass immigration to the United States from the European periphery.

2.2 The Economics of Migrant Self-Selection

Theoretical foundations for the economic analysis of migrant self-selection are laid out by Borjas (1987). Ac-

cording to his modification of the Roy (1951) model, the nature of self-selection into migration is determined

by the relative returns to skill in the sending and receiving economies. If the returns to skill in the receiving

country are higher relative to those in one source country than in another, emigrants from the latter are

predicted to be more strongly positively self-selected than those from the former on the basis of skill. In

most studies of migrant self-selection, the relative returns to skill are proxied by the relative inequality of

the income distributions of each country.6 When focusing on relative inequality, positive self-selection of

migrants is predicted to be induced when the income distribution of the sending country is less unequal

than that of the destination country. Conversely, when the sending country’s income distribution is more

unequal than that of the destination country, migrants are predicted to be negatively self-selected. Borjas

(1987), for example, uses this framework to argue that the deteriorating performance of successive cohorts

of immigrants in the latter half of the twentieth century (as measured by their earnings and integration into

the American labor market) can be explained by the fact these immigrants have increasingly originated in

very unequal countries, and are thus of lower quality.7

This so-called relative inequality model has met with mixed empirical success. Chiquiar and Hanson6We are grateful to Timothy Hatton for pointing out this distinction.7Interestingly, similar arguments were made to explain the labor market performance of the “new” immigrants from the

southern and eastern European periphery as compared to that of immigrants from northern and western Europe in the early20th century.

9

(2005) find that immigrants from Mexico, in which the income distribution is very unequal, are not negatively

self-selected on the basis of earnings, skill, or education. Other recent empirical studies have also found

evidence of positive self-selection into migration, usually with respect to skill or education, regardless of the

relative inequality of income distributions (Feliciano, 2005; Gould and Moav, 2010; Grogger and Hanson,

2011). These findings are rationalized by the presence of migration costs and borrowing constraints that

disproportionately inhibit migration by those in the lower tail of the income distribution (Chiquiar and

Hanson, 2005; Chiswick, 1999; Mishara, 2007). Therefore, regardless of the nature of self-selection of those

wishing to migrate, only those of higher quality are able to overcome liquidity constraints and actually do

so, generating more positive self-selection. However, Ibarraran and Lubotsky (2007) and Fernández-Huertas

Moraga (2011) dispute this finding. Several explanations have been offered to reconcile these results. Borger

(2009), McKenzie and Rapoport (2010), Spitzer (2013), and Wegge (1998) argue that the direction of self-

selection is indelibly tied to the strength of the potential migrant’s social network. Stronger social connections

in the destination country enable individuals who would otherwise be unable to overcome liquidity constraints

to migration to do so, resulting in a weakening of the distortive effects of migration costs and borrowing

constraints on the Roy model effects. Belot and Hatton (2012) and Fernández-Huertas Moraga (2013)

attempt to reconcile these disparate theories. Belot and Hatton (2012) show that once poverty constraints

are accounted for, patterns of self-selection appear to correspond to the predictions of the relative inequality

model. Fernández-Huertas Moraga (2013) finds that a combination of the three explanations is required

to fully account for differences in the pattern of self-selection between urban Mexico (whence migrants are

negatively self-selected) and rural Mexico (whence they are positively self-selected).

A different explanation for the composition of migratory flows is given by development economists, who

have recently examined the role of risk in the migration decision, which is highlighted by Harris and Todaro

(1970). Bryan, Chowdhury, and Mobarak (2014) find evidence that risk aversion prevents rural-to-urban

migration in developing countries. Those with greater wealth, who would be better able to bear the risk,

would therefore be more likely to migrate, generating positive self-selection.

Overall, despite a vast literature studying the issue of migrant self-selection in both modern and historical

contexts, a consensus on its nature and its causes and mechanisms remains elusive. Findings of different

directions of self-selection in different studies make external generalizations difficult.

10

2.2.1 Issues of Data Availability

Empirical disagreement regarding the direction and causes of migrant self-selection can be partially attributed

to data limitations that prevent or restrict comparisons between migrants and the population at risk for

migration in the migrants’ economy of origin. In the absence of other data, many studies have relied on

aggregate data of the place of origin to study migrant self-selection (Bohlin and Eurenius, 2010; Hatton and

Williamson, 1998; Lowell, 1987; Runblom and Norman, 1976). This approach is generally used when micro

data are unavailable, for example, when only aggregate statistics on the volume of migration between two

countries are available. In this approach, self-selection into migration is studied by comparing migration

patterns across regions. If, for example, the rate of migration from higher income areas is greater than

that from other areas, then the conclusion is drawn that migrants are positively self-selected on the basis of

income. The approach is confounded, however, if self-selection also occurs within regions. Returning to the

previous example, migrants from areas with higher average income may be poorer than non-migrants from

the same area and are thus properly understood to be negatively self-selected on the basis of income.

Ideally, micro data would be used, pinning down the types of migrants and permitting the comparison of

individuals within a specific (possibly sub-national) source population to one another. However, samples in

which prospective migrants’ quality is observed prior to migration are rare (Akee, 2010). Instead, scholars

using micro data to study modern migration are often forced to rely on data collected after the migration

has taken place (Chiquiar and Hanson, 2005; McKenzie and Rapoport, 2010). Except for certain indicators,

measures collected after the migrants have been in the receiving economy for some time are likely to be

contaminated by the experience in the destination. For example, occupations of immigrants may change

depending on the labor market conditions of the receiving country.8 Most studies in which such data are

available focus only on very small migration flows, such as from Pacific island nations to the United States

and New Zealand (Akee, 2010; McKenzie, Gibson, and Stillman, 2010), or from Finland to Sweden (Rooth

and Saarela, 2007).

Even if pre-migration data on migrants are available, a comparison between migrants and non-migrants

requires data on the distribution of productive characteristics of the population of origin. Without data on

the population of origin, it cannot be determined whether migrants are positively or negatively self-selected.

For example, individuals with low education in an absolute sense may in fact be highly educated relative to

8For example, Perlmann (2000) shows that the share of laborers and manufacturing workers among Jewish immigrantsduring the Age of Mass Migration was much higher than in the population of origin. Occupation-based self-selection is possible,but Perlmann (2000) argues that many migrants may simply have changed occupations on arrival. Ferrie (1997) also raises theissue that many immigrants tend to work in different occupations before and after migration.

11

their population of origin, but this cannot be determined without data on the source population. Chiquiar

and Hanson (2005) show that failing to take this issue into account can lead to spurious conclusions regarding

the nature of self-selection.

Fernández-Huertas Moraga’s (2013) study of Mexican migrants to the United States is one of the few

that is able to overcome these constraints. He compares survey data on migrants to that on non-migrants

from surveys conducted before migration occurs, and is able to compare the two routs on the basis of

wages, unemployment rates, and labor market participation. However, Fernández-Huertas Moraga’s (2013)

study, like nearly all studies of modern migration, suffer from a problem of dual selection. That is, two

sample selection processes operate to determine the composition of modern migratory flows: the process

that determines whether migrants find it optimal to migrate (which the literature on migrant self-selection

is interested in understanding), and the selection caused by restrictive immigration policy; the latter process

generally obscures the former, and comparisons between migrants and non-migrants do not reveal the nature

of migrant self-selection.

2.2.2 Advantages of Historical Data

Historical data make it possible to overcome some of these hurdles. First, the problem of dual selection

generally does not apply. Specifically, prior to the literacy restrictions imposed in 1917, migration to the

United States from Europe was almost entirely unrestricted. Even after the literacy test was imposed, Goldin

(1994) hypothesizes that migration from Europe to the United States was not significantly restricted until

the quotas imposed in 1924. Thus, migrants in the Age of Mass Migration were selected only by the process

about which we wish to learn: that which determines whether individuals find it optimal to migrate. It is

therefore possible to identify cleanly the self-selection of migrants at the source. Second, historical data often

provide access to micro data that are unavailable to researchers in modern contexts. For example, unlike

most modern data, historical data are not subject to confidentiality restrictions. Moreover, much modern

migration is undocumented; the lack of significant legal restrictions to migration in our study period made

it unnecessary to enter the United States illicitly, ensuring that most migration was documented.

Several studies have used historical data to overcome these data limitations. Abramitzky, Boustan,

and Eriksson (2013) exploit the availability of tax data in Norway to study self-selection on the basis of

wealth into migration to the United States during the Age of Mass Migration. Unfortunately, Abramitzky,

Boustan, and Eriksson (2013) are forced, by issues of data availability, to rely primarily on a binary indicator

of whether a household owned any taxable assets, and are, in their analysis of international migration, unable

12

to further disaggregate wealth. The issue of coarseness also arises when occupational data are used, such as

by Abramitzky, Boustan, and Eriksson (2012), Biavaschi and Elsner (2013), and Wegge (1999, 2002). Such

data generally require that individuals be characterized as either skilled or unskilled, masking much useful

variation in the quality of prospective migrants. It is also possible to rank occupations by median income

(Abramitzky, Boustan, and Eriksson, 2012; Biavaschi and Elsner, 2013), but no variation within occupations

is recovered. For instance, this approach cannot differentiate between poor and wealthy farmers, who would

have had vastly different living standards. Indirect inference can also be made from the post-migration

outcomes of immigrants and their children. For instance, Ferrie and Mokyr (1994) find higher rates of

entrepreneurship among immigrants than natives, suggesting positive self-selection. Moreover, Abramitzky,

Boustan, and Eriksson (2014) find that immigrants from some European countries in this period hold higher-

paid jobs than natives on arrival, suggesting that they may also have been positively self-selected.

In most historical literature, however, even coarse data on traditional economic indicators are generally

unavailable.9 Instead, two alternative methods of measuring migrant quality relative to the population at

risk for migration are common in the historical literature. Mokyr (1983) and Mokyr and Ó Gráda (1982),

for instance, use age heaping, based on individual age reports, to infer the numeracy of Irish immigrants to

the United States. Stolz and Baten (2012) perform a similar analysis, comparing migrants from a number

of countries to census records.

Finally, even when all of these constraints can be overcome, it is generally only possible to evaluate the

self-selection of migrants on a national level; that is, migrants are generally classified only by their country of

origin. Analysis of self-selection on the national level, however, may obscure self-selection at the local level

as a result of composition effects across sub-national entities, leading to incorrect conclusions regarding the

true nature of self-selection. Fernández-Huertas Moraga’s (2013) study of modern Mexican immigration to

the United States is, in part, an exception to this restriction. While he does not distinguish between different

geographic places of origin within Mexico in determining self-selection, he does distinguish between migrants

from urban and rural areas in Mexico, finding that migrants from each sector are self-selected differently.

Pooling the sample shows evidence of negative self-selection, obscuring differences in incentives and ability

to migrate among different sectors of the population. As we show in the present research, Italy is a case in

point, with migrants from North and South Italy exhibiting different patterns of self-selection.

9Wegge (2002) also collects data on the wealth of migrants, but systematic misreporting of this figure due to restrictionson the expatriation of cash, together with the lack of comparable data for non-migrants, makes it difficult to draw conclusionsregarding self-selection with respect to wealth.

13

2.2.3 Stature as a Measure of Pre-Migration Living Standards

The use of stature as a measure of economic capability and productivity is grounded in a large literature.

Fogel (1986, 1994), Fogel, Engerman, and Trussell (1982), Komlos and Meermann (2007), and Steckel (1995)

summarize the vast literature establishing a relationship between adult stature and the standard of living

experienced by a population in youth. With the genetic variation in height across individuals averaging

out in comparisons of large groups to one another (Eveleth and Tanner, 1976; Frisancho, 1993; Martorell

and Habicht, 1986; Silventoinen, 2003; Steckel, 1995),10 the average stature of a population represents the

difference between its gross nutrition in youth (principally, its caloric intake) and contemporaneous demands

on nutrition, such as labor and disease. Thus, in addition to being correlated with traditional measures of

the standard of living, such as income or GDP per capita, stature captures additional facets of welfare such

as health and consumption (Steckel, 1995). The variation in stature is also informative about the degree

of inequality in the population in the consumption of inputs to stature (such as food, health, and leisure)

(Komlos, 1985, 1990; Komlos and Baten, 2004; Steckel, 1995)—a feature that Stolz and Baten (2012) exploit

in order to measure inequality when other data are lacking.

Stature is not only correlated with inputs to individual productivity. In essence, stature is a compos-

ite measure of human economic capability—an amalgamation of all factors that ultimately determine an

individual’s standard of living. Thus stature reflects overall quality and economic capability in two ways.

Individuals facing better conditions in childhood (e.g., more food availability, less disease, less hard work)

will both become taller as adults and will also develop superior cognitive skills (Case and Paxson, 2008).

These individuals tend to become more educated than their peers (Case, Paxson, and Islam, 2009), and to

earn higher wages (Lundborg, Nystedt, and Rooth, 2009; Persico, Postlewaite, and Silverman, 2004) and en-

ter into higher-skilled work (Komlos, 1990). Height might also reflect unobserved ability through an indirect

channel: if the provision of a better childhood environment, which would make children taller, is correlated

with parents’ characteristics, such as ambition and resourcefulness, then taller children are more likely to

have inherited such productive characteristics from their parents. Furthermore, for certain occupations,

there are returns to strength, which is correlated with height (Bodenhorn, Guinnane, and Mroz, 2013).

Stature data can therefore be used to address several of the shortcomings of previous studies of migrant

self-selection by overcoming many of the data limitations that they have faced. While other measures

of pre-migration welfare, such as occupation and wealth, have their own advantages,11 the resistance of

10The lack of a genetic difference in adult heights is particularly true when the two groups are from the same place of origin,as is the case in the present research.

11For example, occupational status is measured with less idiosyncratic noise, and thus can be used in cases in which only

14

stature to contamination by post-migration events, its correlation with unobserved ability, education, health,

consumption, and pre-migration welfare, and the availability of finely measured stature data make it an

attractive tool for the study of migrant selectivity. Applying this approach specifically to Italian migration

also makes it possible to study self-selection from sub-national regions due to the availability of geographically

disaggregated and finely measured data on the stature of the Italian population of the time. The historical

coverage of the data remove the dual selection issue.

Stature has been used by several scholars to study migrant self-selection. Crimmins et al. (2005) examine

the self-selection of Mexican migrants to the United States in the modern context. However, the lack of

geographically disaggregated data and the confounds raised by the dual selection of migrants in modern

data limit the generalizability of the conclusions. Kosack and Ward (2013) expand this approach, analyzing

Mexican immigration to the United States in the early 20th century. Unfortunately, they are unable to

compare the stature of migrants to that of a representative sample of Mexicans as no such sample is known

to exist for their study period. Instead the average stature of migrants is compared to that of volunteer

soldiers and passport applicants. As Bodenhorn, Guinnane, and Mroz (2013) point out, however, both of

these comparison samples are likely to suffer from sample-select biases. Thus, it is impossible to determine

whether the finding that Mexican migrants were taller than soldiers and shorter than passport applicants

is an indication of the self-selection of migrants, of the comparison samples in question, or of some mixture

thereof. Humphries and Leunig (2009) study the location choices of mid-nineteenth-century British seamen

based on height. The scope of conclusions that can be drawn from this study are very limited in their

generalizability to the self-selection of an entire population into international migration.

Our study thus improves upon previous attempts to understand self-selection into migration on a number

of fronts. First, we use an easily and finely measured variable that is known to reflect living standards and

other facets of quality, and whose measurement is not affected by the decision to migrate. Second, we compare

migrants to data on the population of origin that are virtually free of self-selection. Third, our comparison is

disaggregated to the province-birth cohort level, enabling us to study self-selection within small population

bins, as well as the variation in self-selection across time and space—all while remaining cognizant of the

different origin populations of the migrants. Finally, our focus on the Age of Mass Migration allows us to

cleanly ascribe observed self-selection to individuals’ migration decisions, rather than to a combination of

these decisions and restrictive policies.

small samples are possible. It is also directly informative on skill and human capital. In contrast, stature requires large samplesin order to eliminate idiosyncratic differences between individuals.

15

2.3 Self-Selection of Italian Migrants in the Age of Mass Migration

Although there have been a number of attempts to determine the nature of self-selection of Italians migrating

to the United States during the Age of Mass Migration, a clear answer has eluded researchers (Gomellini

and Ó Gráda, 2013). In all cases, the difficulties in studying migrant self-selection discussed above apply.

Arguments of negative self-selection are advanced by Betrán and Pons (2004), who find that skill premia

were falling in Italy and rising in the United States during the Age of Mass Migration. These trends indicate

that unskilled laborers were disproportionately overrepresented in emigration, leading to a relative scarcity

of unskilled labor in Italy. Stolz and Baten (2012) also present evidence of negative self-selection, finding

that age heaping among Italian migrants was greater than among the origin population, suggesting negative

self-selection on the basis of numeracy. Giffoni and Gomellini (2013), studying the relationship between

migration and school dropout rates, support this view, at least partially, arguing that they find no evidence

of positive self-selection of Italy. Arguments for positive selection, however, are advanced by Gomellini and

Ó Gráda (2013), who point out that south Italian immigrants were more likely to be literate than their origin

populations. Notably, the latter study makes a comparison of migrants to their region of origin, while the

former compares immigrants to the entire country.

Anthropometric measures have also been used in this context. Danubio, Amicone, and Vargiu (2005)

sample citizenship petitions filed by Italian immigrants in Massachusetts and find an average height greater

than that reported by A’Hearn (2003) and Federico (2003), and computed by A’Hearn, Peracchi, and Vecchi

(2009) and A’Hearn and Vecchi (2011) for the population of Italy. Gomellini and Ó Gráda (2013) interpret

this result as suggesting positive selection into migration. The present research builds on this strategy by

disaggregating the scope of analysis to the sub-national level of Italy and by using data collected prior to

migration, thus eliminating the possibility of post-migration contamination of stature through continued

growth.

3 Data

The data set used in this paper is novel in two ways. First, it makes use of the stature data in the Ellis

Island arrival records, which we discuss in further detail below. Second, it links Italian migrants to their

places of origin with a great deal of disaggregation. In this section we discuss the collection of our data in

further detail, provide summary statistics, and test whether our sample is representative of the population

of migrants.

16

3.1 Data Sources

The primary data sources for our analysis are the province-birth cohort-level Italian stature distributions

computed from military conscription records by A’Hearn, Peracchi, and Vecchi (2009) and A’Hearn and

Vecchi (2011), in addition to the Ellis Island arrival records database. These two sources are discussed

immediately below. We also supplement these data with population and literacy data from the Direzione

Generale della Statistica e del Lavoro (1912) and the Ministero di Agricoltura, Industria e Commercio (1915,

1925).

3.1.1 Italian Stature Data

Analysis of self-selection of any kind based on stature requires a comparison sample that is known to ac-

curately represent the population at risk for migration, or at least to represent non-migrants as a whole,

without further self-selection. We are fortunate that such data exist in the Italian case. As a comparison

sample for our migrants, we use height data compiled as a part of the Italian military conscription process.12

During the period in question, all Italian males, regardless of physical condition, were required to present

themselves for a medical examination, during which their heights were measured and recorded (A’Hearn,

Peracchi, and Vecchi, 2009). As these data are the product of nearly the full male population of Italy

(Cole, 1995), they are representative of the population as a whole. In particular, these data (as corrected

by A’Hearn, Peracchi, and Vecchi, 2009; A’Hearn and Vecchi, 2011) are unlikely to suffer from issues of

self-selection that are problematic in the historical heights literature.

We acquired two sets of data based on the conscription data, one from A’Hearn, Peracchi, and Vecchi

(2009) and the other from A’Hearn and Vecchi (2011). The A’Hearn and Vecchi (2011) data contain the

raw means and standard deviations of the height distributions of each province (except Caserta) and birth

cohort from 1855–1910, as well as these values standardized to their age-20 values. We refer to the latter

data as the “Unsmoothed Age 20” series. Examples of the time series of means and standard deviations of

the unsmoothed age-20 distributions are presented in Figure 4. The distributions of age-20 stature may not

be suitable for comparison to those of migrant stature due to the possibility of growth after age 20. Although

Beard and Blaser (2002) and Frisancho (1993) show that modern populations reach terminal height by age

20, the same need not be true of our study population. Indeed, a number of studies (A’Hearn, Peracchi, and

Vecchi, 2009; Fogel, Engerman, and Trussell, 1982; Frisancho, 1993) discuss the potential for malnutrition to

both reduce final adult height and to delay the onset of the adolescent growth spurt (AGS), leading growth

12For a detailed description of the data and their origin and collection, see A’Hearn, Peracchi, and Vecchi (2009).

17

to continue into the early 20s. Similarly, A’Hearn, Peracchi, and Vecchi (2009) report that the delayed AGS

may be responsible for a decline in the standard deviation of height in a population with age.

Unfortunately, the delayed AGS is a poorly quantified anthropometric phenomenon. We were not able

to find any literature quantifying the effect of nutrition on the rate of growth over the lifespan, and thus

we have only a limited understanding of the bias introduced by using the age-20 distributions as a basis for

comparison. In particular, we do not know to what extent the bias (i.e., the continued growth after age

20) depends on average height at age 20. What we do know, however, is that shorter cohorts are likely to

continue growing further into their twenties. That is, the age 20 distributions may be an image of height

that is earlier in the growth process for shorter populations than for taller populations. This issue would

thus create a mechanical bias toward finding stronger positive self-selection among shorter populations.13

We therefore take advantage of computations performed by A’Hearn, Peracchi, and Vecchi (2009). The

primary computations of these authors resulted in the lines in Figure 4 labeled “Smoothed Age 20,” which

represent the time-smoothed age-20-corrected means and standard deviations of stature. They also adjust

these distributions for continued growth to age 22, with the results represented in the series labeled “Smoothed

Age 22.” These distributions are based on changes in the timing of measurement over the lifespan by the

Italian military,14 but are, for the most part, out-of-sample projections performed by A’Hearn, Peracchi, and

Vecchi (2009). Nonetheless, the growth that these adjusted height distributions depict relative to the age 20

distributions constitutes the most rigorous possible analysis of post-age-20 growth for the population under

analysis. However, the smoothed age-22 distributions eliminate potentially valuable within-province variation

over time. We therefore compute an unsmoothed age 22 distribution, labeled “Implied Age 22” in Figure

4, by adjusting the unsmoothed age-20 means by the province-birth cohort-specific difference between the

smoothed age-20 and smoothed age-22 means. We perform a similar operation on the standard deviations

of the distributions, which are similarly smoothed by A’Hearn, Peracchi, and Vecchi (2009) and not by

A’Hearn and Vecchi (2011). By performing this correction, we produce province-birth cohort-specific height

distributions normalized to age 22, at which the risk of further growth would have become negligible even in

malnourished populations. We therefore consider these distributions to be the best available representations

13We have replicated the results of this paper using the age 20 distributions for comparison. The magnitude of self-selectionthat we find is much stronger than our main results in this paper.

14Specifically, there were variations in the age of measurement by the Italian military induced by the military calling updifferent birth cohorts at different ages. A’Hearn, Peracchi, and Vecchi (2009) report that the vast majority of birth cohortsare measured at age 20, but that for institutional reasons, some were measured as early as age 18, and others as late as age22. Based on this variation, A’Hearn, Peracchi, and Vecchi (2009) compute the average stature at age 22 for each province andbirth cohort by extrapolating from the age 20 distributions that they observe using the differences in the stature observed incohorts measured at different ages.

18

of the average adult height of each birth-cohort and province.15 The time trend in the average height of the

Italian population is depicted in Figure 5.

It may be the case, however, that some smoothing of these moments is necessary. There is likely very

little sampling error in the moments, as they come from nearly the entire male population; but there may

be error in the reporting of ages at Ellis Island that leads us to assign passengers to the wrong birth cohort,

and thus to the wrong comparison distribution. Some smoothing of the moments over time may therefore be

necessary. We therefore compute for each province a kernel regression of each moment against the birth year,

thus providing a smoothed version of the moments of the distributions.16 The smoothed moments are also

presented in Figure 4, and are labeled “Our Smoothed Age 22.” Comparison to these distributions produces

results that are not appreciably different from those driven by comparisons to unsmoothed distributions

except in a small number of cases noted below.

3.1.2 Ship Manifests

Our information on the stature and other personal characteristics of migrants is taken from the Ellis Island

arrival records database, which includes information on nearly all passengers who passed through the Port of

New York from 1897 to 1925.17 This database is composed of passenger manifests deposited at Ellis Island, of

which Figure 6 presents an example. Some of the information on these manifests is already transcribed, while

the rest is available in handwritten form on the scanned manifests. These manifests were completed upon

embarkation by the steamship companies transporting the passengers to Ellis Island, and were primarily

intended to fulfill two purposes. First, they were used to maintain statistics on the number of immigrants of

each gender and nationality entering the United States. Second, they were part of an effort to ensure that

immigrants who might become a public charge, who were ill, or were otherwise undesirable (for instance, by

being anarchists or polygamists), were prevented from entering the United States (Bureau of Immigration

and Naturalization, 1909). Steamship companies were therefore required to assert that they had examined

all passengers, and to affirm that they did not violate any of these restrictions. Beginning in late 1906,

with the passage of the Immigration Act of 1906 (US Congress, 1907), passenger manifests were required to

include a physical description of the passenger, of which height was a component.

15We have also produced all of the results presented below with the age 20 distributions as the point of comparison. Allresults are stronger with the age 20 distributions than with the age 22 distributions.

16We compute our own smoother in order to provide a province-specific average over time. We do not use the smoothedmeans computed by A’Hearn, Peracchi, and Vecchi (2009) because they are not simply averages over time, but are insteadaffected by the temporal trend in other provinces.

17The first five years during which Ellis Island was in operation (1892–1897) are only partially covered for two reasons. First,Ellis Island at this time operated in conjunction with the older Castle Garden facility, where some immigrants were processed.Second, an 1897 fire at Ellis Island destroyed many records that were stored there.

19

Figure 8 presents the time series of arrivals according to both the Ellis Island data base and the official

immigration statistics of the United States (Ferenczi and Wilcox, 1929, Tables 2 and 3, pp. 384–393). The

former exceeds the latter for two reasons. First, it includes both immigrants and individuals entering tem-

porarily, while the official immigration statistics only include people entering with the intention of remaining

permanently. We also include in Figure 8 the Ellis Island statistics deflated by the number of individuals in

our sample who report being first-time arrivals—a proxy for the number of actual immigrants. Second, the

Ellis Island data include individuals who purchased passage but never embarked;18 the official immigration

statistics include only actual arrivals. The official statistics may also include individuals not included in the

Ellis Island data, as the Port of New York was not the only place of entry for Italian migrants,19 though it

was the location of the bulk of arrivals.

We acquired from the Statue of Liberty-Ellis Island Foundation (SOLEIF) a subset of this database,

consisting of the transcribed information of the roughly 4.8 million individuals passing through Ellis Island

in this period who either reported their ethnicity as Italian, north Italian or south Italian, or whose country

of origin was Italy. We restricted the sample to those arriving in 1907 or later so as to consider only those

whose heights would have been recorded under the new law.20 This restriction left approximately 2.8 million

passengers in the sample. Next, we geocoded the passengers’ reported last place of residence using an

algorithm outlined in Appendix A.21 As we discuss in Appendix A, a variety of tests and exercises show

that this algorithm is remarkably accurate for the individuals who can be matched: the rate of false matches

may be below five percent. Moreover, as shown in Figure 7, the correlation of average provincial heights

of men aged 22–65 recorded at Ellis Island and average provincial heights reported by the Italian military

(as adjusted by us and by A’Hearn and Vecchi, 2011) is 0.72. In section 3.3, we analyze whether there are

differences between individuals who were matched, and those who were not. We formally explore the possible

effects of incorrect geolocation on our results in section 7.2.

We then sampled approximately 50,000 passengers arriving after 1907, for whom we transcribed infor-

mation from the original manifests that was not already transcribed by the SOLEIF.22 The data that we18We thank Drew Keeling for pointing this issue out to us. Our present sample includes individuals who did not embark; we

will adjust the sample in future transcription.19Secondary ports of arrival, such as Boston, New Orleans, and Philadelphia also received substantial migratory flows from

Italy; but all of these together amounted to a small share of the total.20There were also a small number of passengers who reported a place of residence in Italy but an ethnicity other than Italian,

north Italian, or south Italian. We also omit these individuals from consideration.21A possibly more appropriate indicator of an immigrant’s origin would be the place of birth; however, unlike the last place

of residence, this information is not available in digital form. If internal migration was common in Italy at the time, there wouldbe differences between the two locations that could lead to incorrect assignment of individuals to comparison distributions. Infuture work, we will transcribe the place of birth of a sample of migrants and compare them to the last place of residence inorder to determine the extent of possible error generated by using the last place of residence instead of the place of birth inorder to select a comparison distribution for each individual.

22We transcribed a simple random sample of households (identified by the ordering of individuals on the manifests and by a

20

received in digital form (indicated by the dashed lines in Figure 6a) included the passenger’s name, marital

status, age, date of arrival, ethnicity, nationality, and last place of residence. We transcribed the answers

to four additional questions asked regarding the migrant (indicated by dashed thick solid lines in Figure

6b): whether he had paid for his own passage, and if not, who had paid for the passage; whom he would be

joining in the United States; whether he had ever been in the United States before; and his height.23

3.2 Summary Statistics

Figure 9 depicts the arrivals of Italian passengers in the entire 1907–1925 period, disaggregated by the

province to which they were matched by our geolocation algorithm. A striking pattern is evident: first,

southerners were much more likely to migrate to the United States than their northern counterparts. In

particular, the rates of emigration in the regions of Sicily and Abruzzo are over 12 percent, while those in

Emilia Romagna were nearer to two percent. Moreover, southerners represented a much larger proportion

of all Italian passengers traveling to the United States than did northerners. Eleven provinces of southern

Italy, and none in northern Italy, were the origins of more than 50,000 geolocated passengers each. Moreover,

nearly all provinces from which fewer than 5,000 geolocated passengers originated are located in the north.

In total, 82 percent of passengers in our geolocated sample are matched to a southern province.24

We restrict our sample to individuals aged 22–65 who could be matched to a province of origin by our

geolocation algorithm.25 We make this age restriction and retain it throughout the paper, as this is the range

of ages over which we can be confident that terminal height has been achieved, but rapid shrinking has not

begun.26 Moreover, we see a peculiar pattern in the age distribution of male migrants, which is illustrated in

Figure 10. As is typical of the Age of Mass Migration, the density of the age distribution is greatest in the

early twenties. There is, however, a large dip in the distribution between ages 18 and 21, a trend that we do

not observe among Italian women, and which is not present, for instance, among Russian Jewish immigrants

(Spitzer, 2013, 2014). We believe that this dip, which corresponds to the age of military service, may be

common last name) and not of individuals. Thus, an individual traveling with one companion was twice as likely to be sampledas an individual traveling alone. Of all passengers between 1907 and 1925, nearly 75 percent traveled alone, and 94 percenttraveled in groups of three or less. All further discussions are therefore corrected for this sampling technique through the useof appropriate weights.

23We are very grateful to Roy Mill for providing us with access to his dEntry transcription system, and for devotingconsiderable time and effort to making it compatible with the requirements of this project.

24This figure falls to 81% when individuals from Caserta, for which we do not have population stature information, aredropped.

25We examine whether our algorithm induced sample selection bias in section 3.3.26Cline et al. (1989) show that shrinking begins essentially as soon as final height is attained, but accelerates with age. In

any event, changing the end point of our sample in terms of age will not have large effects on our results, as there are relativelyfew older immigrants as compared to younger ones. The distribution of ages in our sample is illustrated in Figure 10. In thefew cases in which our results are qualitatively affected by reducing the terminal age of our sample to 40 (which is a moreconventional terminal age Silventoinen, 2003), we describe the difference.

21

attributable to the restrictions on legal emigration for males in this age range (Cole, 1995). We therefore

suspect that migrants in this age range are self-selected differently from their fellow countrymen emigrating

at a later age.27

We present summary statistics in Table 1. Column (1) presents summary statistics for all men and

women in our geolocated sample, restricting attention to the fields for which no transcription was necessary.

Consistent with official statistics (Ferenczi and Wilcox, 1929), we see that the immigrants are overwhelmingly

male—more than 75 percent of our sample. Moreover, approximately 70 percent of passengers in our sample

reported being married. Approximately 85 percent are matched to a province in southern Italy, as defined

by the Bureau of Immigration. Columns (2) and (3) present these statistics for males and females separately.

Female passengers are, on average, older than male passengers, approximately as likely to be married, and

very slightly less likely to be from southern Italy.

In columns (4)–(9), we restrict attention to the sample of individuals for whom we transcribed additional

information. Column (4) presents the information for all transcribed individuals, while column (5) presents

it for females and column (6) presents it for males. The already-digitized information for each group is

similar to that for the untranscribed sample. Based on our transcription, we classified any passenger listing

any person whom they would be joining in the United States as having some connection, and any individual

who reported joining an immediate family member in the United States (i.e., a sibling, parent, child, or

spouse) as having an immediate family connection in the United States. Over 95 percent of male and female

migrants report that they have some connection (“Any Conn.”), but males were far less likely to report that

this connection was an immediate family member (i.e., a parent, sibling, child, or spouse), with only 32

percent falling in this group (“Imm. Fam. Conn.”= 1) as compared to nearly 74 percent of women. Similar

differences are apparent in the fraction of men and women reporting having been in the United States

before (“Repeater”). More than 40 percent of men reported that they had been in the United States before,

compared to only 16 percent of women. A gender difference is also apparent in whether the passenger had

paid for himself, with 90 percent of men paying their own passage and only 66 percent of women doing so.

Moreover, an unusual relationship exists between the heights of men and women, the distributions of which

are presented graphically in Figure 11. In particular, female passengers were much taller relative to male

passengers than is commonly the case in modern populations (Gaulin and Boster, 1985). This relationship

is discussed in more detail in Appendix B. As discussed in this Appendix, however, we find no reason to

believe, based on this relationship, that there are systematic issues of accuracy in our data. However, given

27Comparisons to the age 20 distributions of province and birth cohort height indicate that those migrating at the ages of20 and 21 are negatively self-selected.

22

that we do not have data on the distributions of stature for women in Italy, we exclude women from our

analysis.

In column (7), we eliminate from the sample any passenger who indicated that he had been in the United

States before. We make this restriction primarily for two reasons. First, the process of self-selection into

return migration is not well understood (though it has recently received scholarly attention: Abramitzky,

Boustan, and Eriksson, 2012, 2014; Bandiera, Rasul, and Viarengo, 2013; Crimmins et al., 2005; Ward,

2013). Distinguishing between first-time and return migration prevents our sample from being contaminated

by some other form of self-selection (i.e., into return migration) and prevents us from counting the same

passengers more than once. Second, these passengers may have arrived in the United States before completing

their physical growth, and would thus have grown differently (Boas, 1911, 1920; Gravlee, Bernard, and

Leonard, 2003; Kress, 2007; Sparks and Jantz, 2002, 2003). Therefore, our benchmark sample will be that

summarized in column (7)—males aged 22–65 who reported never having been in the United States. The

remaining migrants after this deletion are younger than the repeat passengers, less likely to be married or

to have an immediate family connection, and very slightly shorter. All of the results discussed below are

stronger when these repeat passengers are included.

Next, we summarize the geographic distribution of male heights graphically. Figure 12 presents the

average male heights of each province (based on A’Hearn, Peracchi, and Vecchi, 2009 and A’Hearn and

Vecchi, 2011) weighted by our passenger counts, as well as the average heights of male passengers in our

sample from each province. The average military heights exhibit a strong pattern, with the tallest provinces

in the north, the shortest in the south, and the middling provinces in the center. We see a similar trend in

the heights of migrants, with the tallest originating in the north, and the shortest in the south. Column (7)

of table 1 also shows that the average heights of male migrants in our sample was 163.80 cm.28

Finally, we study separately two samples that allow us to break down the analysis by time period. As

is evident from Figure 8, World War I was a massive disruption to trans-Atlantic migration. Moreover,

in 1917, the United States enacted the literacy test, requiring that all adults entering the United States

demonstrate literacy. Both of these events fundamentally changed international migration, and there is

reason to believe that post-1917 passengers might be substantially different from pre-1917 passengers, and

that pre-War migrants may have differed from the post-War migrants. We therefore split the sample into

pre-1917 (exclusive) and post-1917 (inclusive) subsamples, which are summarized in columns (8) and (9),

28By contrast, the average American soldier (who may have been negatively self-selected) born in the 1860s was 171 cm tall(Zehetmayer, 2011, Figure 1, p. 320).

23

respectively, of Table 1.29 Most striking is the large difference in stature between the two periods: passengers

in the post-1917 are more than one centimeter taller on average. They are also more than eight percentage

points more likely to have an immediate family connection, six percentage points less likely to be married,

and four percentage points more likely to have paid for their own passage as compared to those arriving

before 1917.

3.3 Representativeness of the Geolocated Sample

Before beginning our primary analysis, we examine whether our geolocation algorithm has produced a

representative sample for our analysis. First, we estimate a number of regressions of the form

yi = β0 + β1Gi + εi,

where yi is some individual characteristic of interest, and Gi is an indicator equal to one if individual i is

successfully matched to a province by our algorithm, and zero otherwise. The coefficient β1 tests whether

there is a difference in the mean of each characteristic between the geolocated and non-geolocated groups.

Table 2 presents estimates of β1 for a variety of individual characteristics of interest and for a variety of

samples. One division of the data is based on the recorded ethnicity of passengers traveling from Italy.

Beginning in 1903, the passenger manifests were required to include the ethnicity (“Race or People”) of

immigrants (Perlmann, 2001; Weil, 2000). North Italians and south Italians were officially considered to be

two separate ethnicities. The instructions for clerks completing the passenger manifests placed the dividing

line between north and south Italy at the southern extreme of the basin of the River Po.30 Nevertheless,

compliance with the official definitions of the ethnicities appears to have been lax, and some passengers were

still recorded as simply Italian, without further disaggregation. We refer to these passengers as “General

Italians.” Figure 13 also depicts the division of Italy into North and South by the Bureau of Immigration and

Naturalization. This field provides information on the probable geographic origin of individuals independent

29We could also split the sample into pre-1914 (inclusive) and post-1919 (inclusive) subsamples in order to omit the WorldWar I years, but given the fall in the quantity of migration during this period, there is essentially no difference between thisapproach and ours.

30Specifically, the manifest defined north Italians as[t]he people who are native to the basin of the River Po in northern Italy (i.e., compartments of Piedmont,Lombardy, Venetia, and Emilia) and their descendants, whether residing in Italy, Switzerland, Austria-Hungary,or any other country . . . . Most of these people speak a Gallic dialect of the Italian language.

South Italians were defined as[t]he people who are native to that portion of Italy south of the basin of the River Po (i.e., compartments ofLiguria, Tuscany, the Marches, Umbria, Rome, the Abruzzi and Molise, Campania, Apulia, Basilicata, Calabria,Sicily, and Sardinia) and their descendants . . . .

24

of our algorithm. We use it in Appendix A in order to test the accuracy of our algorithm, and here in

order to determine if our algorithm led to imbalances between matched and unmatched individuals within

geographic regions.

We begin in column (1) of Table 2 by studying all males aged 22–65 in our 1907–1925 group of passengers.

Our analysis indicates that south Italians are significantly overrepresented in our geolocated sample (66.5

percent of the geolocated sample, as opposed to 58.6 percent of the non-geolocated sample) while north

Italians are slightly underrepresented and general Italians are significantly underrepresented. This under-

representation of general Italians is likely due to the fact that Italians traveling through non-Italian ports

were less likely to be assigned a north/south ethnicity, and their locations are likely to have been recorded

with less accuracy due to a lack of familiarity by clerks in other countries with Italian geography and

spelling.31 We also find that those in our geolocated sample are on average 0.08 years younger than the

non-geolocated. In addition, the average birth year is 0.35 years later. There is no statistically significant

difference in marriage rates between groups.

As we are interested in observing differential self-selection patterns across provinces, we must verify that

the sample is balanced at the provincial level, which can be approximated (without using our geolocation

algorithm) by ethnicity. In columns (2)–(4) of Table 2, we break down the the sample used in column (1) by

the ethnicity of migrants. There exist statistically significant differences in the probability of being married

between the geolocated and non-geolocated groups for each ethnicity, but these are small, likely reflecting

the large sample sizes as much as any actual differences. Similarly, differences in age and birth year exist,

but are also small. Moreover, differences in age and birth year are not particularly troubling, as all of our

analyses condition on birth year by comparing the height data of migrants to the averages of their birth

cohort, and height is constant with age in our range.

Next, we perform the same exercises restricting attention to the transcribed sample. Columns (5)–(8)

report the difference in the means of various characteristics between transcribed males aged 22–65 who were

matched to a province by our algorithm, and those who were not matched. Statistically and economically

significant differences persist in matching rates across ethnicities. However, there are no statistically sig-

nificant differences in age, birth year, or marital status between groups, even within ethnicities. We also

compare matched and unmatched individuals on the basis of transcribed information. We find no statisti-

cally significant differences between matched and unmatched individuals on the basis of transcribed data,

except among northerners with respect to the measures of social network status and whether the passengers

31For example, the modal departure port for ships characterizing all Italian passengers as simply Italian was Cherbourg,France, while the modal departure port for ships decomposing all Italian passengers by ethnicity was Naples.

25

paid for their own passage. In particular, matched individuals are four percentage points less likely to have

any connection in the United States, nine percentage points less likely to have an immediate family connec-

tion, and four percentage points more likely to have paid for their own passage than unmatched individuals.

These differences have no implications for our main results (which are based only on height). They could

potentially have implications for our results regarding the mechanisms driving self-selection.

The fact that our main results are based on stature makes it particularly important to test the repre-

sentativeness of our sample on this front. In particular, there are two potential dangers that we face in

terms of the balancedness of our sample. First, as we compare the heights of migrants to those of the Italian

population, it is important to ensure that our sample of migrants is representative of all migrants, rather

than being biased upward or downward in height by our algorithm. Although not statistically significant,

Table 2 shows that matched north Italians were 0.375 centimeters taller than their unmatched fellows, while

south Italians were 0.009 centimeters shorter. If these difference represent a small but non-spurious bias,

they would bias our estimates against our baseline results of more positive self-selection in the south. Figure

14 shows this result in greater detail: among passengers of all ethnicities, the probability of being matched

is all but constant over height. Within the separate ethnicities, there is more noise among the rare heights,

but essentially the same conclusion follows.

Second, since we test whether, within a province, migrants are taller or shorter than their source popula-

tion, it is important to ensure that, conditional on the province of origin and its mean height, our matched

sample is not taller or shorter than the unmatched group. Finally, since we test whether the trends in

self-selection differ across cohorts and provinces or different mean heights, we must ensure that there are no

differences in the differences between the heights of the matched and unmatched individuals across cohorts

of different average stature. However, determining the province to which an individual belongs requires a

successful match, which we do not have for the matched individuals. We must therefore find some way

of associating unmatched individuals with provinces and birth cohorts independently of the geolocation

algorithm.

To this end, we use the following procedure, which takes advantage of the fact that Italian surnames

are useful indicators of geographic origins (Guglielmino and De Silvestri, 1995). First, for each surname,

we determine the modal province to which individuals with that surname who could be geolocated were

assigned. Then, for the purposes of this exercise only, we assign all individuals to the modal province for

their surname. We then use their (known) birth year to assign them to a province and birth cohort, from

which a mean height for each individual’s province and birth cohort is determined. The rationale behind

26

this exercise is the following: family names can be used to group migrants into bins that are clustered across

space. If geolocated individuals are randomly drawn from within each province and birth cohort, then we

expect that the height distributions of matched and unmatched passengers would be the same within each

surname bin. Mapping passengers to a province predicted by their surnames brings us as close as possible

to comparing height distributions within provinces and birth cohorts and makes it possible to test whether

the matched-unmatched gap changes systematically with the height of the province and birth cohort.

We first use the results of this procedure to estimate the regression equation

zijt = β0 + β1µjt + β2Gijt + β3Gijtµjt + εijt, (1)

where zijt represents the standardized (by the surname-implied province-birth cohort mean and standard

deviation) height of individual i, who is matched (by this procedure) to province j and birth cohort t, with

mean height µjt (normalized to have mean zero), and where Gijt is an indicator equal to one if individual

i was matched to a province by our geolocation algorithm and to zero otherwise. We present the results

of this regression in Table 3. There are two coefficients of interest. First, β2 indicates whether there is, at

the average, a systematic difference in standardized heights between the geolocated and the non-geolocated.

While we find that the geolocated individuals are, on average, taller than the non-geolocated conditional

on the mean height of their place of origin, this difference is statistically insignificant. This difference is

somewhat concerning, as it would tend to spuriously generate our findings in section 4.3. However, it must

be kept in mind that the unmatched comprise less than 15 percent of migrants, which would make any

sample selection biases induced by our algorithm small. Second, β3 indicates whether there is a systematic

difference in the difference between the matched and unmatched groups between provinces and cohorts of

different average heights. We find that this coefficient is positive, but that again, it is not statistically

significant. Moreover, the positive sign works against our results in section 4.4. Thus, the true differential

patterns are stronger than those that we measure.

We also seek to verify that these findings are not driven by the linearity assumptions of equation (1).

We therefore regress non-parametrically, in Figure 15a, zijt on µjt for each of the two groups of Gijt, and

present an estimate of the difference between the two curves in Figure 15b.32 Throughout the range of µjt,

the confidence band includes zero. Moreover, the nature of the difference, as above, is such that it would

work against our differential results, except at the upper extreme of mean heights, at which the data are

32The confidence bands in the graph are 95 percent point wise confidence intervals. We thank Anand Krishnamurthy forhelpful discussions on this topic.

27

sparse. Thus, on the whole, these balancing tests do not show any compelling evidence of differences in

heights or differential differences in heights between the matched and unmatched groups, nor compelling

reason to believe that our results will be driven by imbalances in our geolocation algorithm, However, the

point estimates show that we cannot rule out sample-selection bias entirely, and we will thus discuss the

potential consequences of sample-selection bias below.

4 The Nature and Degree of Self-Selection

We are now equipped to begin examining the nature and degree of self-selection of Italian migrants. We

first lay out a formal framework for our analysis. We then study migrants as compared to all of Italy

before disaggregating the analysis to compare these migrants to their provinces of origin. We then study

geograp